| The detection and localization of underwater Marine organisms have very important research significance and value for the rational development and utilization of Marine resources.The underwater environment is complex and special,and it is easy to be disturbed by the environment,so the traditional target detection method cannot meet the detection requirements.Compared with traditional target detection methods,the deeplearning-based target detection method has high detection accuracy,but poor detection effect on underwater targets and frequent missed and false detection.Therefore,this paper proposes two underwater target detection methods based on deep learning,and the main research contents are as follows:A two-stage underwater target detection method based on cross-scale feature hopping fusion and dual attention network is proposed.By changing the conventional feature extraction method and designing cross-scale feature jump fusion,multi-scale target feature is extracted and target feature information is enhanced.Due to many underwater interference factors and complex background,the feature information is similar.Therefore,by constructing a dual attention mechanism,this method gives more attention to the feature information of the target channel and the spatial location information,so as to improve the discriminability of the target and the background,and to accurately locate the target.Finally,in order to obtain more difficult samples,this method adopts a new sampling method to improve the overall sample quality through adaptive balancing of positive and negative samples,so as to efficiently train positive samples and achieve the best training effect through fast network convergence to improve the performance of the detector.Experiments on Pascal VOC,URPC2019 and Wild Fish standard data sets showed that m AP reached 80.9%,85.3% and 96.9%,respectively.The detection time of single image was 0.57 s,and its detection performance was better than that of mainstream detection methods.A single-stage underwater target detection method based on Ghost Net and cross-path feature aggregation is proposed.Firstly,the feature representation layer is reconstructed to reduce certain pooling layers and add small target information of shallow feature map to improve the detection rate of small targets.Secondly,according to the characteristics of underwater targets,by improving the feature pyramid structure,a cross-path feature aggregation network is proposed to reduce underwater noise interference and enhance the classification ability of the network.In the underwater scene,the target detection rate is low.By building an efficient attention mechanism,the location information can be embedded in the channel and the detector can be re-encoded with Transformer to strengthen the target features and location information and improve the network regression ability and detection rate.In order to make the network lighter,lightweight Ghost Net is used instead of Resunit to reduce the number of network parameters.The m AP of this method on Pascal VOC,URPC2019 and Wild Fish standard data set reached82.3%,89.1% and 97.3%,respectively.The detection frame rate of this method was81.5FPS and the detection time of a single image was 12.3ms,which was superior to the advanced detection method.Compared with cross-scale feature jump fusion and dual attention underwater target detection methods,m AP improved by 1.4%,3.8% and 0.4%,respectively. |
